- Featured
- Coursework
2026 · Computer vision · Responsible AI · Coursework · COMM042 Computer Vision, University of Exeter
Campus Safety Object Detection
Built a YOLOv8n object detector for person, bicycle, car, stop-sign and traffic-light classes using public COCO images and campus-specific photographs.
- Python
- YOLOv8
- OpenCV
- PyTorch
- Computer Vision
Headline
5 classes
Person, bicycle, car, stop-sign, traffic-light
On this page
01 · Problem
Problem
Object detection is easy to demo and hard to trust. Models that look accurate on curated test sets fail on occlusion, small objects and edge-case lighting, which matter most in safety-critical settings.
02 · Approach
Approach
A YOLOv8n detector was trained on public COCO images combined with campus-specific photographs, and evaluated with mAP, IoU and confidence-threshold analysis, including targeted occlusion and small-object test cases.
Figure
How it fits together
Illustrative layout, not detections from the project
03 · Results
Results
A working detection pipeline for the five classes, evaluated against mAP, IoU and confidence-threshold behaviour.
04 · Limitations
Limitations
Surveillance risk, dataset bias, and the false-positive / false-negative trade-off were assessed explicitly as part of the academic project, alongside the technical evaluation.